01 / Discovery
Tell us about the task, permitted tools, success criteria, security boundaries, and data rights. We agree on a narrow scope before any customer data is transferred.
Proposed service · Discovery first
Training environments built around a task you can actually check.
We are exploring bespoke pilot work for teams building and evaluating AI agents. A pilot would pair a bounded task environment with objective checks and a replay-backed evaluation report. It is custom work, not a turnkey RL gym or a promise of model-training gains.
Ask about a scoped pilot →Opens your email app. No information is sent until you choose to send it. Please do not include secrets or customer records in the first note.
Where a pilot might fit
Candidate domains, not finished products: simulated infrastructure and DevOps incidents, e-commerce operations, and code review. The first discovery call tests whether an outcome can be checked without subjective reward shortcuts.
A proposed pilot shape
02 / Scoped build
For a proposed pilot, we would define a resettable task, action and observation contract, isolated state, objective reward checks, and adversarial cases. Deliverables and timeline depend on discovery.
03 / Evidence
The proposed handoff includes a versioned evaluation set, traces, repeatable results and failure analysis. ConeReplay can support trace and replay work; it is not a hosted training platform.
What we can show today
An offline, synthetic DevOps-incident reference in the ConeReplay repository demonstrates resettable seeded episodes, bounded actions, deterministic reward checks and a replay-backed before/after fixture evaluation. Its tests and container runner passed in CI on Python 3.11 and 3.12. This is fixture evidence only: it does not measure live-model learning, customer performance, or a deployed customer environment.
Reference CI run (repository access may be required) ↗Start with a conversation
Best fit: a lab, platform team, or agent team with a specific task, a way to check success, and an owner for scope and data handling. Tell us the domain and what a useful pilot would prove. We quote only after discovery; no price, delivery time or result is guaranteed here.
Email about a pilot →Your email goes to r@persooninc.com when you send it. We use your message and reply address to respond to this inquiry. Review the privacy draft before sharing personal data. No customer traces or credentials by email.